This article reports the panel discussion on agentic AI as an emerging paradigm of computing. It took place at IEEE 2026 International Congress on Intelligent and Service Oriented Systems Engineering (CISOSE 2026) on 28th July 2026. The panel addressed key questions about Agentic AI from a wide range of perspectives, which include the characteristics of this emerging paradigm of computing, its social impact on trustworthiness computing and education, the current state of art, the directions of future development and technical challenges.

Modern cities increasingly rely on heterogeneous data sources and intelligent systems to improve urban management and quality of life. Among the many dimensions of urban intelligence, air quality monitoring and prediction stand out as a critical service given the well-documented impact of atmospheric pollutants on public health. This paper presents a framework that repurposes existing traffic camera infrastructures, combined with meteorological data, to forecast pollutant concentrations without requiring extensive historical air quality records. A comparative evaluation of six deep learning architectures, namely Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), Convolutional LSTM (CNNLSTM), LSTM with Attention, CNNLSTM with Attention, and Transformer, is conducted across two major Spanish cities, Madrid and Bilbao. Each architecture is evaluated over 10 independent runs with different random seeds, and statistical significance is assessed through Friedman tests followed by pairwise Wilcoxon signedrank tests with Holm-Bonferroni correction. The CNNLSTM and LSTM with Attention architectures are statistically the best-performing models overall, with no significant difference between them (mean RMSE of 5.754 and 5.672, respectively; p = 0.846). However, no single architecture dominates all scenarios: the Transformer achieves the best results for SO2, PM2.5, and PM10 in Bilbao, whereas the LSTM and LSTM with Attention form the statistically best group for O3. These findings highlight that pollutant-specific model selection, rather than a one-size-fits-all approach, is essential for operational short-term urban air quality forecasting.

The analysis of the privacy properties of Privacy-Preserving Ads APIs is an area of research that has received strong interest from academics, industry, and regulators. Despite this interest, the empirical study of these methods is severely hindered by the lack of publicly available data. Reliable empirical analysis of the privacy properties of an API, in fact, requires access to a dataset consisting of realistic API outputs for a large collection of users; however, privacy concerns prevent the general release of such data to the public. In this work, we address this problem by developing a novel methodology to construct synthetic API outputs that are simultaneously realistic enough to enable accurate study and provide strong privacy protections. We focus on one of the Privacy-Preserving Ads APIs: the Topics API, part of Google Chrome's Privacy Sandbox, which enables interest-based advertising without relying on third-party cookies. We developed a methodology to generate a differentially-private dataset that closely matches the re-identification risk properties of the real Topics API data. The use of differential privacy provides strong theoretical bounds on the leakage of private user information from this release. Our methodology is based on first computing a large number of differentially-private statistics describing how output API traces evolve over time. Then, we design a parameterized distribution over sequences of API traces and optimize its parameters so that they closely match the statistics obtained. Finally, we create the synthetic data by drawing from this distribution. Our work is complemented by an open-source release of the anonymized dataset obtained by this methodology. We hope this will enable external researchers to analyze the API in-depth and replicate prior and future work on a realistic large-scale dataset. We believe that this work will contribute to fostering transparency regarding the privacy properties of Privacy-Preserving Ads APIs.

With the increasing adoption of AI in safety-critical applications within urban environments, the interpretability of these systems is paramount. This study explores the application of Explainable Artificial Intelligence (XAI) techniques to enhance transparency in audio-based detection of emergency vehicle sirens, a crucial component in urban sound management. Adopting methods such as SHAP (SHapley Additive exPlanations) values, Permutation Feature Importance, and model-specific feature scores, this research identifies key audio features, including mid-frequency spectral contrasts and targeted chroma components, which significantly help in distinguishing siren sounds among urban noise. The study examines various machine learning models, identifying K-Nearest Neighbors (KNN) and XGBoost as top performers; KNN excelled in class-specific precision, while XGBoost demonstrated strong cross-class discrimination. The findings highlight the potential of XAI in improving both accuracy and accountability for sound detection systems in safety-critical urban applications, advancing the deployment of transparent AI within smart city infrastructures.

With increasing urbanization, efficient urban traffic management is a critical challenge that requires smarter and more adaptable systems. This paper introduces a self-learning algorithm designed to enhance the adaptability and effectiveness of vehicle detection models using urban camera infrastructures. By leveraging these ubiquitous devices, the study aims to capture and analyze real-time traffic data, a task traditionally limited by the need for extensive manual data labeling and the limitations of pre-trained models under varying urban conditions. Our self-learning algorithm addresses these challenges by reducing reliance on manual labeling and enabling continuous model adaptation to new conditions without direct human intervention. Implemented in the dynamic urban environment of the city of Madrid, Spain, this study evaluates the algorithm’s capacity to enhance vehicle detection, considering a diverse range of vehicle types. The core of the algorithm comprises an iterative self-training process that refines model performance using both labeled and unlabeled data, thus progressively enhancing detection accuracy. Our findings reveal significant improvements in the ability of the model to accurately identify and classify vehicles, highlighting the potential of self-learning algorithms in urban traffic management.

This paper presents a novel Multi-DL Fuzzy Approach aimed at performing image recognition in the development of a real-time traffic alert system, addressing the problem of traffic congestion and related incidents. Traditional monitoring by road operators predominantly relies on fixed location cameras, yielding limited and sometimes ambiguous information. This study proposes leveraging Twitter (now known as ‘X’) as a more comprehensive data source alongside employing fuzzy techniques with Deep Learning (DL) neural networks such as CNN, VGG16, and Xception to analyze and classify traffic images. The innovative integration of these technologies augments the precision in categorizing varying traffic conditions, namely fluid and dense traffic, accidents and fires. Thus, this proposal mitigates the ambiguities prevalent in traffic image interpretation, and reduces the dependency on static data sources. The proposed models showed improved results by combining information from the DL models, elevating accuracy from 84% in crisp classification to 90% utilizing fuzzy information.

The music industry is now more complex and competitive than ever before. In recent years, the search for collaborations with other artists has become a common strategy for musicians to maintain their presence in the sector. Besides, existing music streaming services such as Spotify have exposed large data feeds that can be used to develop innovative services within the realm of music. In this context, the present work introduces PRESTO, a novel recommendation system to suggest musicians for new collaborations with other artists by means of an ensemble of Graph Neural Networks. The system is fed with a heterogeneous graph representing the time evolution and the stationary aspects of a musician’s career. Finally, the proposal has been evaluated with a dataset comprising more than 200,000 artists, with an average F1 score above 0.75.

When analyzing the energy consumption of a car, several studies estimate that the air-conditioning system can account for up to 10% of such consumption. Other studies have found that the cabin temperature may exceed 60°C when it is exposed to direct sunlight, but it can be reduced by more than 15°C if the vehicle is covered. Hence, this chapter introduces TICKET, an intelligent parking space allocation system for greater energy efficiency. TICKET informs drivers of which parking space they should occupy within the parking lot so that when the driver returns to collect the vehicle, the cabin temperature is as low as possible. To do so, the system considers multiple environmental factors such as the sunlight movement or the weather conditions in the target parking lot and also combines predictive algorithms like hidden Markov models for demand estimation. Finally, TICKET has been tested in a real parking-lot premises with promising results.

This paper explores the limitations faced by current solutions for selecting quasi-identifying attributes in the context of Privacy-Preserving Data Publishing (PPDP). PPDP stipulates that any published personal data should not be linkable to other available data sources in a manner that could potentially lead to individual re-identification or compromise sensitive data. The state-of-the-art methods for selecting quasi-identifying attributes commonly rely on heuristic evaluations to assess the risk of re-identification associated with each attribute. We hypothesize that these heuristic-based methods could be significantly improved by complementing them with empirical methods capable of quantifying the external linkability of dataset attributes. This empirical layer would enable a fine-tuning of the obfuscation of attributes within the dataset, thereby preventing the unnecessary privatization of attributes beyond potential attackers' reach while ensuring privatization of those easily accessible. For this purpose, we explore recent advancements in identifying semantically related datasets across heterogeneous data sources. Although initially developed for purposes beyond privacy preservation, these methods support our initiative by uncovering potential links with external data and thus providing empirical evidence for the identification of attributes as quasi-identifiers. Finally, we discuss potential pathways to implement this empirical layer in quasi-identifier identification systems.

Sentiment analysis has gained prominence in the digital era, particularly within social media contexts. This study explores the integration of metadata, such as reposts, favorites, or followers, alongside textual information for sentiment analysis. A dataset of posts in Spanish is created, filtering by a set of keywords (a total of 17) of interest to us. A total of 3,962 posts were collected along with their metadata. Then, they have been manually classified as positive, negative, and neutral, and processed. The research studies and compares different models like LightGBM, BERT, and a hybrid model combining BERT outputs with metadata. Evaluation metrics, such as F1-score and precision, were employed, revealing that the hybrid model achieved the highest F1-score of 69.4%, showcasing the effectiveness of merging text comprehension capabilities with contextual metadata. While the incorporation of metadata enhanced overall sentiment classification, precision varied across sentiment classes, with notable improvements in neutral and positive sentiments, over 16% and 8% respectively.

Greenhouses constitute intricate systems where numerous variables play a pivotal role in enhancing crop yields within the framework of intensive agriculture. Consequently, real-time monitoring and visualization of these variables are imperative to strike a balance between resource efficiency and production maximization. Furthermore, the ability to make predictive assessments regarding these variables is essential to avert potential greenhouse disasters. In this article, we introduce an intelligent alert system designed to efficiently oversee agricultural operations within a functioning greenhouse, ultimately bolstering productivity through the optimization of crop growth and energy consumption. This system comprises a web application, GreenhouseGuard, which improves the graphical and statistical representation of data collected by a network of sensors strategically positioned throughout the greenhouse, as well as the forecasts generated from this data. These sensors are strategically located to provide more precise real-time data readings, thereby minimizing error margins. Moreover, GreenhouseGuard offers diverse data visualization options and forecasts of greenhouse variables to enable in-depth analysis of the acquired information. Consequently, this alert system empowers greenhouse managers to proactively address abnormal situations that may jeopardize their crop yields.

Time-series analysis plays a crucial role in extracting meaningful patterns from sequential data, serving to gain insights into temporal trends. Specifically, the burgeoning urbanization trend accentuates the urgency of traffic forecasting in modern cities due to its socio-economic and environmental impact. While Deep Learning techniques, notably Long Short-Term Memory (LSTM) networks, have shown promise in predicting traffic, they often overlook contextual factors like weather and holidays. To address this challenge, this paper proposes a hybrid model combining LSTM with categorical and continuous data to forecast traffic volume on the Interstate 94 American highway. Incorporating weather, temporal patterns, and holidays, this study explores the performance of a model that includes contextual factors against standalone LSTM. The results show superior predictive accuracy for the proposed hybrid model. SHapley Additive exPlanations (SHAP) analysis reveals the influence of diverse features, emphasizing the significance of contextual attributes in enhancing traffic prediction.

Traffic congestion is a significant problem in many cities, leading to delays, increased fuel consumption, and environmental pollution. In recent years, optimization algorithms have been used to improve traffic control systems and reduce congestion. This paper proposes an optimization algorithm based on genetic algorithms to optimize traffic signal controllers in Cádiz, Spain, using the Simulation of Urban MObility (SUMO) traffic simulator. The proposed optimization algorithm is a promising solution for the optimization of traffic control systems, which can lead to more efficient traffic flow, reduced congestion, and improved environmental conditions.

A well-known numerical bottleneck in the differentially-private stochastic gradient descent (DP-SGD) algorithm is the computation of the gradient norm for each example in a large input batch. When the loss function in DP-SGD is consists of an intermediate linear operation, existing methods in the literature have proposed decompositions of gradients that are amenable to fast norm computations. In this paper, we present a framework that generalizes the above approach to arbitrary (possibly nonlinear) intermediate operations. Moreover, we show that for certain operations, such as fully-connected and embedding layer computations, further improvements to the runtime and storage costs of existing decompositions can be deduced using certain components of our framework. Finally, preliminary numerical experiments are given to demonstrate the substantial effects of the aforementioned improvements.

We study differentially private mechanisms for sharing training data in machine learning settings. Our goal is to enable learning of an accurate predictive model while protecting the privacy of each user's label. Previous work established privacy guarantees that assumed the features are public and given exogenously, a setting known as label differential privacy. In some scenarios, this can be a strong assumption that removes the interplay between features and labels from the privacy analysis. We relax this approach and instead assume the features are drawn from a distribution that depends on the private labels. We first show that simply adding noise to the label, as in previous work, can lead to an arbitrarily weak privacy guarantee, and also present methods for estimating this privacy loss from data. We then present a new mechanism that replaces some training examples with synthetically generated data, and show that our mechanism has a much better privacy-utility tradeoff if the synthetic data is realistic, in a certain quantifiable sense. Finally, we empirically validate our theoretical analysis.

The management of road traffic incidents is a problem faced by governments in many countries. Normally, road operators have the infrastructure in place to monitor such incidents, albeit in a reactive manner. In Spain, there are traffic cameras on major roads to check for possible incidents, however, incident notification is slow and not automated. As an alternative, this paper proposes a system for automatic real-time traffic alerts. Thus, 1,500 camera images from the Dirección General de Tráfico (DGT) deployed on the main Spanish roads are analyzed in real time every 4 minutes. These images are not preprocessed, they have different qualities and are also affected by weather conditions such as fog, rain, sun reflections, etc. The system uses several Deep Learning classification models trained on a well-known dataset of traffic images including flowing traffic, dense traffic, accidents and fires. These models are used to classify the DGT images in real time, with satisfactory initial results, detecting both flowing traffic and dense traffic.

Tourism has become a very active ecosystem to deploy solutions based on Information and Communication Technologies. Indeed, it is now possible to analyse the mobility behaviour of tourists in great detail. However, current solutions aimed at anticipating tourist flows usually follow a limited approach based on the local (e.g., to predict the next landmark to visit) or regional (e.g., to predict the incoming number of tourists in a city) level. This paper states a novel approach to solve the problem of tourist inflow forecasting on a broader nationwide scale by defining it as an edge prediction task. To do so, we model the tourist mobility of a country as a graph which fuses heterogeneous tourism data obtained from multiple sources related to the country's mobility and infrastructure features. Then, as a major contribution, an ensemble of Graph Neural Networks are fed with the graph models to provide the final prediction. The proposed solution has been tested in Spain showing a F1 score higher than 0.7.

We consider the problem of Learning from Label Proportions (LLP), a weakly supervised classification setup where instances are grouped into "bags", and only the frequency of class labels at each bag is available. Albeit, the objective of the learner is to achieve low task loss at an individual instance level. Here we propose Easyllp: a flexible and simple-to-implement debiasing approach based on aggregate labels, which operates on arbitrary loss functions. Our technique allows us to accurately estimate the expected loss of an arbitrary model at an individual level. We showcase the flexibility of our approach by applying it to popular learning frameworks, like Empirical Risk Minimization (ERM) and Stochastic Gradient Descent (SGD) with provable guarantees on instance level performance. More concretely, we exhibit a variance reduction technique that makes the quality of LLP learning deteriorate only by a factor of k (k being bag size) in both ERM and SGD setups, as compared to full supervision. Finally, we validate our theoretical results on multiple datasets demonstrating our algorithm performs as well or better than previous LLP approaches in spite of its simplicity.